Accurate detection of structural anomalies in bridges is crucial due to the variety of deteriorations that can compromise their safety. Traditional visual inspection methods are subjective and can miss incipient damage. This study proposes a methodology to evaluate the performance of a model based on unsupervised learning for bridge anomaly detection. Real acceleration data from a Warren truss-type steel bridge are used, which are pre-processed through standardization (mean zero and standard deviation one) and normalization (scaling the data between −1 and 1) and are divided into sequences of 5 s with a 200 Hz sampling rate to adequately capture temporal variations. An MLP Autoencoder with multiple hidden layers is implemented, using cross validation (KFold), Adam optimizer and mean square error (MSE) loss for training. The performance evaluation includes visualizations with PCA (Principal Component Analysis) and UMAP (Uniform Manifold Approximation and Projection), clustering with K-means, and analysis of reconstruction errors, calculating metrics such as precision, recall and F1-score, and presenting the confusion matrix. The results showed validation errors between 0.029 and 0.047, indicating moderate consistency of the model. Principal component analysis and UMAP revealed clear structures in the data, while K-Means clustering identified five distinct clusters. The model achieved an overall accuracy of 62.50%, with 76% for the majority class. This methodology provides a basis for the automated characterization of structural vibration patterns, contributing to the development of more sensitive and efficient monitoring systems for the early detection of bridge anomalies.